Negative BOLD-fMRI Signals in Large Cerebral Veins

Abstract

Reductions in blood oxygenation level dependent (BOLD)-functional magnetic resonance imaging (fMRI) signals below baseline levels have been observed under several conditions as negative activation in task-activation studies or anticorrelation in resting-state experiments. Converging evidence suggests that negative BOLD signals (NBSs) can generally be explained by local reductions in neural activity. Here, we report on NBSs that accompany hemodynamic changes in regions devoid of neural tissue. The NBSs were investigated with high-resolution studies of the visual cortex (VC) at 7T. Task-activation studies were performed to localize a task-positive area in the VC. During rest, robust negative correlation with the task-positive region was observed in focal regions near the ventricles and dispersed throughout the VC. Both positive and NBSs were dependent on behavioral condition. Comparison with high-resolution structural images showed that negatively correlated regions overlapped with larger pial and ependymal veins near sulcal and ventricular cerebrospinal fluid (CSF). Results from multiecho fMRI showed that NBSs were consistent with increases in local blood volume. These findings confirm theoretical predictions that tie neural activity to blood volume increases, which tend to counteract positive fMRI signal changes associated with increased blood oxygenation. This effect may be more salient in high-resolution studies, in which positive and NBS may be more often spatially distinct.

Introduction

An intriguing aspect of blood oxygenation level dependent (BOLD)-functional magnetic resonance imaging (fMRI) experiments is that certain brain regions show reduced signal level with increasing task demands (Shmuel et al, 2002, 2006; Raichle et al, 2001; Allison et al, 2000). A striking example of this is the signal in the visual system during strong visual stimulation, in regions surrounding the activated cortex (Shmuel et al, 2002, 2006; Smith et al, 2000). Initially, this phenomenon was attributed to a hemodynamic effect resulting from the so-called ‘blood stealing’ phenomenon (Harel et al, 2002; Shmuel et al, 2002). However, electrical recordings suggest that this negative BOLD signal (NBS) results from neural inhibition (Devor et al, 2007; Shmuel et al, 2006). Another example of negative BOLD activation is found in a cluster of brain regions often referred to as the default mode system, which shows reduced BOLD signal, positron emission tomograph blood flow, and CMRO₂ during increasing cognitive demands and goal directed action (Raichle et al, 2001). In the absence of task demands (also called the resting state), default mode system shows negative correlation with the dorsal attention system (Fox et al, 2005; Greicius et al, 2003). These findings are consistent with the notion that negative BOLD signals (NBSs) represents a reduction in neuronal activity. During both rest and task execution, NBSs may manifest as a negative correlation between brain signals, reflecting an inhibitory interaction between regions involved in competitive processing. Finally, negative BOLD effects can be observed in the stimulus-induced hemodynamic response, in terms of an initial dip and a poststimulus undershoot. Those transient negative BOLD responses have been attributed to hemodynamic mechanisms rather than neuronal inhibition (Buxton et al, 2004; Uludag, 2010).

In this work, we report on a novel pure hemodynamic mechanism producing NBS, as suggested by several recent fMRI studies. During a combined motor/visual stimulation protocol, van der Zwaag et al (2009) found substantial BOLD signal reductions in periventricular areas apparently devoid of neural tissue. In the absence of stimuli during REM sleep (Hong et al, 2009) and resting with the eyes closed during states of varying vigilance (Olbrich et al, 2009; Moosmann et al, 2003), similar regions also appear to negatively correlate with visual areas. These findings raise the possibility that at least in certain brain regions, NBS may not reflect a reduction in local neural activity but rather a hemodynamic phenomenon. If confirmed, this may impact the understanding of the BOLD contrast mechanism and the interpretation of BOLD-fMRI studies in general.

The goal of this study was to further investigate NBSs and their potential nonneuronal contribution. For this purpose, BOLD-fMRI studies of spontaneous activity in the visual system and periventricular areas were compared with high-resolution anatomical scans obtained at 7T, and contributing sources of NBSs were identified by investigating the underlying contrast mechanism.

Materials and methods

To investigate NBSs and confirm the earlier findings in periventricular areas (Hong et al, 2009; Olbrich et al, 2009; van der Zwaag et al, 2009; Moosmann et al, 2003), we first reanalyzed resting-state single-echo BOLD-fMRI data acquired in previous work (Bianciardi et al, 2009a). Subsequently, to accurately identify the areas with negative signals, we repeated the fMRI protocol of Bianciardi et al (2009a) and also acquired high-resolution anatomical data. Finally, to investigate the origin of negative BOLD, we studied the underlying contrast mechanism by acquiring multiecho echo-planar imaging (EPI) images. In the following, each of these three experiments is discussed in detail. All studies were performed according to a human subject protocol approved by the Institutional Review Board of the National Institutes of Health. All scans were performed at 7T with multichannel receive-only coil elements, sensitivity encoding (SENSE) acceleration in the phase encode direction, and real-time shimming of B₀ fluctuations because of respiration (van Gelderen et al, 2007). For fMRI data, the first 10 images (which also allowed the magnetization to reach a steady state) were used as a reference for coil sensitivity, and then discarded from further analysis.

Experiment 1: Negative Correlations with Spontaneous Activity in the Visual Cortex

Eleven complete data sets from a previous study (Bianciardi et al, 2009a) were analyzed. Two behavioral resting conditions were investigated: eyes open with visual fixation, and eyes closed. A region of interest (ROI) was selected in the visual cortex (VC) from a functional localizer based on a rotating wedge paradigm (see Figure 1). For each condition, which lasted about 6 minutes, multislice gradient-echo-EPI data were acquired from 16 receive coils at 7T (General Electric Medical Systems, Waukesha, WI, USA) using echo time (TE)/repetition time (TR) = 32 milliseconds/3 seconds, voxel size = 1.25 × 1.25 × 2 mm³, 105 time points (see Bianciardi et al, 2009a for further details).

Figure 1

(A) Stimulus used to localize the visual cortex (VC). The flickering (7.5 Hz) B/W wedge (width in polar angle: 24°) performed a full clockwise rotation in the visual field, covering 30 different positions in 90 seconds (see Bianciardi et al, 2009a for additional details). (B) Areas in the brain that respond to the wedge (Figure 1A) across all rotations form region of interest ROI_VC (orange, shown for an example data set: Experiment 1, slice = 19).

Data preprocessing: Slice timing, motion correction and image coregistration were performed as described previously (Bianciardi et al, 2009a). Regression analysis was performed to remove signal variability originating from a number of confounds including: low frequency drifts (third degree polynomials), cardiac and respiratory cycles (eight RETROICOR regressors; Glover et al, 2000), respiratory depth (Birn et al, 2006) and cardiac rate (Shmueli et al, 2007). Effects due to changes of physiologic rates were modeled according to the dual-lagged model (Bianciardi et al, 2009b): the respiration volume (RV) regressor was shifted at time lags −9 and +9 seconds and the cardiac-rate regressor was shifted at time lags −3 and +9 seconds. Global signal regression was not performed, because it may produce spurious anticorrelations (Murphy et al, 2009, Weissenbacher et al, 2009). Finally, fMRI signal fluctuations were converted to percentage signal changes (SC, %) relative to their time average, and low-pass filtering (cutoff frequency = 0.073 Hz) was applied.

Data analysis: For both conditions, multiple regression was performed using a matrix comprised of motion-correction parameters (three rotation angles and three translation displacements), and the averaged time series across ROI_VC (seed time series), modeling spontaneous activity in the VC. The latter regressor was orthogonalized to the motion-correction parameters to separate motion effects from spontaneous activity. The t-test statistical maps were obtained for motion-related effects and spontaneous activity for each subject. An ROI comprising the areas (mainly periventricular regions) anticorrelating (AC) with spontaneous activity in the VC was defined during the eyes-closed condition (t value < −5.7, P < 0.05, Bonferroni corrected, ROI_AC; see Figures 2 and 3).

Figure 2

(A) Sample data set (Experiment 1), left, axial view (slice = 19) and, right, sagittal views (slices = 46, 56) of the spatial distribution (P < 0.05 Bonferroni corrected) of signals correlating (green) and anticorrelating (red) with the seed time series in the visual cortex (VC). Negative correlations cover periventricular areas mainly around the lateral ventricles (especially lining the dorsal part of the inferior and posterior horns), the third ventricle and the fourth ventricle. (B) Average time series in a region of interest in the VC (ROI_VC, green) and in anticorrelating (AC) ROI (ROI_AC, red) for the same data set shown in (A), after (solid lines) and before (dash-dotted lines) physiologic noise correction. The amplitude of spontaneous activity in both (C) ROI_VC and (D) ROI_AC decreased during the fixation condition with respect to resting with the eyes closed (P < 0.001). The amplitude of motion-related effects did not change across conditions in either of the two areas (at a significance level P < 0.05). Error bars display the amplitude average ± standard error (s.e.) over eight subjects. Pearson correlation value between the time series in ROI_AC during resting with the eyes closed and (E) respiration volume (RV) and (F) cardiac-rate regressors shifted at different time lags. The average correlation was not significant for any lag time.

Figure 3

Sample data set (Experiment 2, slices = 31, 32, 33, 38, eyes-closed condition), showing a zoomed view of: (A) the spatial distribution (red/green) of functional magnetic resonance imaging (fMRI) signals anticorrelating and positively correlating with the seed time series in the visual cortex; the anatomy of the lateral ventricles and of the visual cortex in (B) magnitude and (C) phase images (TE = 28.5 milliseconds). Anticorrelated regions overlap with large ependymal and pial veins in proximity of ventricular and sulcal cerebrospinal fluid, respectively. Veins display dark contrast in magnitude anatomical images. In phase images, they display a dipolar-shaped contrast, which depends on the vessel orientation with respect to the static magnetic field and to the acquired slice.

The fluctuation amplitude (% signal change, SC) of spontaneous activity and motion in ROI_VC and ROI_AC was measured as the standard deviation of the signal explained by the spontaneous activity regressor and by the six motion-related regressors in each voxel pertaining to the two ROIs. The average amplitude across each ROI was then computed per subject and the group average (± standard error, s.e.) was displayed (see Figures 2C and 2D).

Spontaneous activity in both the VC and in periventricular areas was defined after removal of instrumental and physiologic noise. For RV and cardiac-rate regressors, we adopted the lags optimized for the VC and for the whole gray matter as reported in previous work (Bianciardi et al, 2009b). As this may be suboptimal for periventricular areas, we computed the Pearson correlation value (r value) between signals in voxels within ROI_AC (after motion correction and before noise correction) and RV and cardiac-rate regressors for the eyes-closed condition. These regressors were shifted by time lags ranging from −21 to 21 seconds (step 3 seconds). The average r value across ROI_AC was then computed for each subject, and the average ± s.e. across subjects was displayed. The latter was compared with previous results obtained for the VC (Bianciardi et al, 2009b).

Experiment 2: Colocalization of Negative BOLD with High-Resolution Anatomical Data

On five subjects (3 males/2 females, age 33 ± 4 years) fMRI data were acquired as in Experiment 1, together with high-resolution multiecho gradient-echo anatomical data. For the latter, we used: TR= 1.9 seconds; TE1, TE2 = 14 milliseconds, 28.5 milliseconds, flip angle = 90°, bandwidth = 62.5 kHz, number of slices = 42, voxel-dim= 0.3125 × 0.3125 × 1 mm³, slice-gap = 1.2 mm; SENSE acceleration rate = 2. For both functional and anatomical scans, an array of 32 receive coils was used, and 42 slices were acquired at matching locations.

Data preprocessing and analysis: The fMRI data were preprocessed and analyzed as in Experiment 1. Both magnitude and phase images were derived from the anatomical data analogously to Duyn et al (2007). On a slice-by-slice basis, phase data were unwrapped with PRELUDE (http://www.fmrib.ox.ac.uk/fsl/) and low-frequency changes were removed with polynomial fitting (model order = 8, Matlab—the Mathworks, Natick, MA, USA—routine). From the comparison of functional and anatomical data, NBSs appeared to overlap with large vessels proximal to cerebrospinal fluid (CSF). Therefore, we aimed at statistically quantifying the spatial overlap of NBSs with vasculature close to CSF. As accurate spatial coregistration between fMRI and high-resolution anatomical images was not always achieved, we plotted the histogram of EPI intensities of voxels pertaining to ROI_AC and ROI_VC in the first fMRI image during eyes closed (after having discarded the first 10 images). For each subject, the area of the histogram was normalized to one. The normalized histogram was then fitted to three Rician-distributed curves (each with area normalized to one). The three Rician curves modeled: (1) a low EPI-intensity component (presumably vessels or partial volume effects between vessels and CSF or tissue); (2) a mid-range EPI-intensity component (mostly gray matter); and (3) a high EPI-intensity component (presumably CSF). Multidimensional nonlinear minimization of a cost function was achieved by the Nelder–Mead method, with initial values for the unknown Rician noncentrality and scale parameters (v, σ) of components 1, 2, and 3, respectively, equal to (0, 20), (75, 50), and (300, 30). The cost function was chosen equal to the residue of the multicomponent fitting to the normalized histogram. For each subject, the ratio of the fitted area of component 1 plus the fitted area of component 3 divided by the fitted area of component 2, R = (A_C1 + A_C3)/A_C2, was quantified in the two ROIs. A two-sample t-test of R was performed across subjects to compare the mean R in ROI_AC versus ROI_VC. This analysis was performed for both data of Experiments 1 and 2 (results were not merged as most of the subjects participated in both experiments).

Experiment 3: Contrast Mechanism Underlying Negative BOLD

On three subjects (three males, age: 34, 34, 23 years) the eyes-closed and functional localizer scans were performed as in Experiment 1. The fMRI data were acquired with multiecho EPI to allow calculation of T*₂ and TE-independent SCs and characterization of their origin.

Data acquisition: For both the eyes-closed condition and the functional localizer, gradient-echo-EPI BOLD-fMRI was performed at 7T (General Electric Medical Systems, Waukesha, WI, USA) using 32 receive-only coil elements and the following parameters: TR = 2 seconds, flip angle (FA) = 75°, number of slices = 20, voxel size = 2.5 × 2.5 × 2 mm³, slice spacing = 2 mm, SENSE acceleration rate = 3. For the eyes-closed condition, multiecho images were acquired: TEs = 12/28.2/44.4 milliseconds were used in odd scans (acquisition times t = (1 × TR, 3 × TR, …), Series A) and TEs = 20.1/36.3/52.5 milliseconds in even scans (acquisition times t = (2 × TR, 4 × TR, …), Series B). Eighty-four images were acquired for each series. For the functional localizer scan, we used TE = 30 milliseconds, 245 time points and otherwise similar parameters.

Data preprocessing: T*₂and S₀ images were computed from multiecho data by linear least squares fitting, assuming S = S₀ exp(–TE/T*₂). For the eyes-closed condition, the same image processing as in Experiment 1 was performed on the single-echo, T*₂ and S₀ images. For both Series A and B, motion correction was computed for the first-echo images and then applied to the second echo, third echo, T*₂ and S₀ images.

Data analysis: For the eyes-closed condition, multiple regression was performed for each echo time as well as the T*₂ and S₀ data. The regression matrix comprised of the averaged time series across ROI_VC (seed time series), modeling spontaneous activity in the VC. The seed time series was computed from the second-echo images, separately for Series A and B.

The t-test statistical maps were obtained for spontaneous activity for each subject. An ROI comprising the (mainly periventricular) areas anticorrelating with spontaneous activity in the VC was defined during the eyes-closed condition in T*₂ images (t value < −5.2, P < 0.05, Bonferroni corrected). The β values from this ROI in T*₂ images and S₀ images were extracted.

Results

In 8 out of 11 subjects (Experiment 1), significant anticorrelation with VC signal was observed in focal spots near the ventricular areas (posterior and anterior horns of the lateral ventricles, third ventricle, and fourth ventricle), consistent with data presented in earlier studies (Hong et al, 2009; Olbrich et al, 2009; van der Zwaag et al, 2009; Moosmann et al, 2003). Additional focal spots of this NBS were observed throughout the occipital cortex. An example is shown in Figure 2. The amplitude of this anticorrelation reached about 1.8% during the eyes-closed condition (Figure 2D). To further investigate the origin of signals in ROI_AC, we also verified if NBSs were time lagged or purely anticorrelated with respect to VC signals. The anticorrelation strength between average VC signals and average NBSs shifted at various time lags showed maximum anticorrelation for zero lag (with interpolation to 150 milliseconds from the original 3-second long time bins, we found that negative signals lagged at most 300 milliseconds the VC signals). A pure antiphase rather than a time delay between signals in ROI_AC and ROI_VC was found for signals both before and after physiological noise correction. The observed anticorrelation was not simply caused by motion artifacts, as significant motion-related (positive as well as negative) effects (P < 0.05 Bonferroni corrected) were limited to few sparsely distributed voxels that showed minimal overlap with the anticorrelated regions (0.17% ± 0.09%, average ± s.e. across subjects for the eyes-closed condition—maps not shown). The dependence of spontaneous activity and motion-related effects on brain region and experimental condition is given in Figures 2C and 2D. These data show that the level of activity in both positively and negatively correlating regions was lower during the fixation condition than during the eyes-closed condition (P < 0.001, unpaired t-test). This was in contrast to the motion-related effects, which were not significantly different between behavioral conditions.

The Pearson correlation value between the time series in ROI_AC during resting with eyes closed and RV and cardiac-rate regressors at various time lags showed a bimodal dependence, see Figures 2E and 2F. The distribution was very similar (with opposite sign) to that previously reported for the VC (Bianciardi et al, 2009b), indicating that the adopted lags optimized for physiologic noise removal in the VC are an optimal choice for periventricular areas as well. Note that the average correlation of spontaneous signal fluctuations in periventricular areas with RV and cardiac-rate regressors did not reach significance for any time lag (r value = 0.19 for P = 0.05 uncorrected for multiple comparisons).

To investigate the source of the anticorrelated activity, fMRI activity maps were compared with the high-resolution anatomical data. Figure 3 shows an example of fMRI negative and positive correlation maps (A), obtained during the eyes-closed condition, together with magnitude (B) and phase (C) anatomical images. Anticorrelated activity was found predominantly in the proximity of the ependymal vascularization of the ventricular system (darker contrast in anatomical magnitude images), most notably near the ependymal lining of the posterior horns of the lateral ventricles (tributaries to the vein of the posterior horn from the medial and the lateral wall of the posterior horn; Ono et al, 1984; Johanson, 1954), and near the glomus and the choroid plexus of the lateral ventricles (choroidal veins). The focal spots of anticorrelation found dispersed throughout the occipital lobe generally colocalized with large pial veins (Serrano et al, 1972).

To quantify the apparent colocalization of large veins and NBS, we investigated the distribution of baseline EPI signal intensity in ROI_AC and in ROI_VC (Figure 4). Across subjects, in ROI_AC with respect to ROI_VC, we found a larger contribution (P < 0.05) of low EPI values (attributed to a large venous blood volume fraction) and high-EPI values (attributed to a large CSF volume fraction) relative to mid-range EPI values (mostly gray matter). A large contribution of vessels in ROI_AC was found for both data in Experiment 1 (on average ± standard error across subjects, R = (A_C1 + A_C3)/A_C2 was equal to 0.42 ± 0.08 in ROI_AC, and 0.21 ± 0.06 in ROI_VC) and in Experiment 2 (R equal to 0.51 ± 0.11 in ROI_AC, 0.13 ± 0.04 in ROI_VC).

Figure 4

Group average (± s.e.) normalized histogram count (black) of functional magnetic resonance imaging (fMRI) baseline signal amplitude between 0 and 450 (a.u.) in voxels displaying positive (A) and negative (B) blood oxygenation level dependent (BOLD) effects during resting with the eyes closed (Experiment 2). Overlayed on the signal amplitudes, we display the fitted Rician-distributed tissue components averaged (± s.e.) across subjects. Note that the fits were performed on single subject data and not directly on the group mean histogram (shown here for display purposes only), because fMRI signal amplitudes are not quantitative values. The multicomponent fit shows a larger contribution of components 1 and 3 relative to component 2 in anticorrelating region of interest, ROI_AC, with respect to the visual cortex ROI, ROI_VC (P < 0.05), in line with a superposition of ROI_AC with large vessels near cerebrospinal fluid. Similar results were also found for data of Experiment 1.

The phenomena described above were reproduced in the multiecho experiment (Experiment 3), which showed significant correlations and anticorrelations at all echo times (Figures 5A and 5B). Furthermore, calculated T*₂ images from the multiecho data mirrored the observations at individual echo times. Interestingly, the polarity of the correlations was reversed in the calculated S₀ maps (Figures 5A and 5B). The anticorrelation between the average signal in ROI_AC in T*₂ and S₀ images (Figures 5C and 5D) was significant (P < 10⁻⁶) for each subject. This phenomenon is quantified in Table 1, which shows the average (± s.e.) β values of the correlation between periventricular areas in T*₂ and S₀ images and the seed time series in the VC. For each subject across voxels, β values were smaller than zero in T*₂ images and greater than zero in S₀ images (P < 0.01).

Figure 5

Maps of correlation (green/red = positive/negative t values) with the seed time series in the visual cortex (VC) (P < 0.05 Bonferroni corrected) for a single subject (Experiment 3) for (A) Series A and (B) Series B, overlaid on a second-echo EPI-image from Series A. In Series A, echo 1/2/3 are acquired at echo times 12/28.2/44.4 milliseconds, in Series B at 20.1/36.3/52.5 milliseconds. Note that negative correlations in periventricular areas persist in T*₂ images and are reversed in sign in S₀ maps. For each time point, the average (± s.e., error bar) T*₂ change relative to baseline (%) across voxels in anticorrelating (AC) region of interest, ROI_AC, versus the average S₀ change relative to baseline (%) across voxels in ROI_AC is shown for (C) Series A and (D) Series B (rows left to right correspond to different subjects participating in Experiment 3). The anticorrelation between T*₂ and S₀ signals at the ROI level was significant for each subject (P < 10⁻⁶).

Table 1

Analysis of the strength and of the sign of the correlation between periventricular areas and the seed time series in the visual cortex (Experiment 3)

	β value a		No. voxels b
	T * 2	S0
Subject 1
Series A c	−0.7 (0.1)	+1.4 (0.3)	22
Series B	−0.6 (0.0)	+1.0 (0.3)	8
Subject 2
Series A	−0.6 (0.0)	+1.1 (0.3)	10
Series B	−0.6 (0.1)	+1.2 (0.4)	22
Subject 3
Series A	−1.6 (0.2)	+1.2 (0.3)	27
Series B	−1.6 (0.1)	+3.6 (0.9)	38

a

β values are averaged across the voxels of periventricular areas and the s.e. across voxels is displayed in parenthesis.

b

No. voxels = number of voxels in periventricular areas. The latter were defined by thresholding t-maps of T*₂ images at P < 0.05 (Bonferroni corrected, t value = 5.2).

c

In Series A, echo 1/2/3 were acquired at echo times 12/28.2/44.4 milliseconds, in Series B at 20.1/36.3/52.5 milliseconds.

To further analyze these findings in periventricular regions, we used a bicompartment model to simulate the contributions of CSF and blood, the two main constituents of this region (see Appendix). Results of this modeling indicate that the opposing changes in T*₂ and S₀ are consistent with blood volume changes. This can be intuitively understood considering that both T₁ and T*₂ are shorter in blood than in CSF, and therefore blood volume increases will reduce both the voxel averaged T₁ and T*₂. Under general fast imaging conditions, apparent T₁ decreases lead to S₀ increases. In contrast, blood T*₂ decreases in the absence of a blood volume increase will reduce the voxel T*₂ but will not affect S₀.

On the basis of this bicompartmental model, we proceeded to estimate the fractional blood volume change required for the observed 1.8% SC at 7T (obtained in periventricular areas on average across subjects in Experiment 1, see Figure 2D, eyes-closed condition). For a vessel with its axis parallel to B₀, the results of calculations are shown in Figure 6A. Starting from a baseline fractional blood volume v _BL between 0.2 and 0.3, which is not unreasonable for the voxel volume occupied by a single big vein, we found blood volume changes between 4.2% and 7.2%.

Figure 6

(A) Vessel with its axis parallel to B₀. Simulation of the expected blood oxygenation level dependent (BOLD)-functional magnetic resonance imaging (fMRI) signal changes (SC, %) in periventricular and perisulcal areas due to increases in blood volume (%), for several baseline blood volume fractions v _BL (see legend) and blood oxygenation Y = 0.6. For this vessel orientation, changes in Y up to 0.7 only minimally affect T*_{2 BLOOD} and do not affect BOLD signal level (the dephasing factor f_DEPH is equal to 1 for any v _BL and Y). A baseline v _BL of 0.05 was only used for 3 and 1.5T simulated data, considering the lower spatial resolution (here, we assumed a 3 × 3 × 3 mm³ voxel volume) usually used at three these field strengths. The acquisition parameters and the assumed T₁ and T*₂ values for the various tissue compartments are explained in Table 2. A SC of −1.8%, found in periventricular and perisulcal areas on average across subjects during resting with the eyes closed (Experiment 1, see Figure 2D), is marked with a horizontal black line. This signal decrease may be generated by a blood volume change of about 4% to 7%, assuming a resting v _BL of 0.2 to 0.3 at 7T. Note that, for the same baseline v _BL and blood volume change, the induced signal decrease is smaller at lower field strengths. (B) Vessel with its axis perpendicular to B₀. The calculated BOLD-fMRI SCs (%) decrease more sharply with the blood volume (%) compared with (A) for baseline Y = 0.6 (dY/Y = 0%). Increases in blood oxygenation counteract this effect (e.g., see curves with dY/Y = 4.4% or 16.7%). This is because f_DEPH decreases with v _BL and increases with Y for vessels with their axis perpendicular to B₀.

Extending this calculation to lower field strengths (1.5 and 3 T), we found that, for the same baseline v _BL and blood volume change, the induced signal decrease is smaller at lower field strengths. For instance, for v _BL = 0.05, a SC of about −1.6%, −1.4%, and −1.0% was induced respectively at 7, 3, and 1.5T by a 30% blood volume increase.

Moreover, for a fixed field strength (e.g., 7T), voxel size, and blood volume change, a larger baseline v _BL (i.e., vessel with larger radius for the same voxel size) induced larger signal decreases. For instance, at 7T, for a voxel of 1.25 × 1.25 × 2 mm³, a SC of −2.15% and −0.25% was induced by a 5% blood volume change of a vessel having a baseline radius, respectively, equal to 500 μm (i.e., baseline v _BL = 0.3) and 100 μm (i.e., baseline v _BL = 0.05) Finally, results of calculations for a vessel with its axis perpendicular to B₀ (Figure 6B) show stronger NBSs compared with vessels in the parallel orientation, in the absence of changes in blood oxygenation or for small oxygenation changes (e.g., dY/Y < 16.7%).

Discussion

In BOLD-fMRI, negative activation in task-evoked studies and anticorrelation in resting-state data have been observed under varying conditions. Although these phenomena may generally represent local neuronal deactivation (Devor et al, 2007, Shmuel et al, 2006, Raichle et al, 2001), the results of the current study suggest that this is not always the case. Here, we show that under particular circumstances, the otherwise robust BOLD contrast mechanism can break down and lead to spurious anticorrelated activity. Using high-resolution fMRI at high field, we show that anticorrelation with the VC is observed in voxels with CSF and large veins and apparently devoid of neural tissue. Such voxels in periventricular regions have been previously shown to exhibit anticorrelation during task performance (van der Zwaag et al, 2009), and rest conditions (Hong et al, 2009; Olbrich et al, 2009; Moosmann et al, 2003). Importantly, the T*₂ and S₀ (i.e., TE-independent) changes in these areas, derived from multiecho data, suggest that blood volume effects in large veins may underlie the observed anticorrelation; the strongest effects would occur in voxels that have high proportions of venous blood and CSF and are downstream from the activated sites (such as periventricular areas and the larger pial veins). Simulations suggest that magnitude of the observed T*₂ and S₀ changes can be plausibly explained with two-compartment model of CSF and venous blood contribution to the voxel signal.

BOLD contamination in large draining veins is a recognized confounding effect in the fMRI literature (Lee et al, 1995; Turner, 2002). Large positive BOLD SCs, lagged by about 3 seconds with respect to the early onset BOLD activation, are reported in large draining veins close to activation areas (Lee et al, 1995), with time delays resulting from transit-time effects as the oxygenated blood passes through the vasculature. Here, we report on a different mechanism in large cerebral veins: first, signals in ROI_AC were not time delayed but purely in antiphase, with respect to those in ROI_VC (a pool of sulcal vessels in the occipital lobe 180° out phase, with respect to an early onset activation was also reported by Lee et al, 1995); second, negative and not positive BOLD signals were found in those large vessels; third, large vessel both close to (pial veins) and remote from (periventricular ependymal veins) the activation area were negatively activated.

Our finding of negative BOLD around large cerebral veins suggests a mismatch between blood volume increases and oxygenation changes in those vessels (see for instance Figure 6B). This invites speculation on the possible causes of such a mismatch. Smaller oxygenation changes in larger versus smaller veins could be explained by downstream oxygen diffusion outside the vasculature or downstream dilution of activation-related blood oxygenation changes (Turner, 2002). For instance, the fractional blood oxygenation change was shown to decrease as a function of distance from activated areas, for example to 25% of its initial value at a distance of 2.5 cm from an activation area of 100 mm² (Turner, 2002; for instance, see in Figure 6B the SCs caused by blood volume changes for an oxygenation change of 4.4% and 16.7%). The reduction in oxygenation change at distal sites might not be matched by an equal reduction in blood volume changes and as a result produce NBSs.

The hypothesis that blood volume changes with little or no change in blood oxygenation cause a decrease in BOLD signals has been previously proposed as an explanation for the poststimulus undershoot (Buxton et al, 2004) and the initial dip (Uludag, 2010) of the BOLD response. The initial dip may be related to blood volume changes in blood vessels with high oxygen concentration (arterioles and arteries; Uludag, 2010); a differential temporal dynamic of blood volume and blood flow changes might be the cause of the BOLD poststimulus undershoot (Buxton et al, 2004). Our findings are in agreement with these predictions that tie neural activity to venous blood volume increases that tend to reduce the fMRI signal. Nevertheless, unlike the BOLD undershoot, we believe that NBSs in large veins are caused by a different relationship between blood volume and oxygenation changes rather than a differential temporal dynamics of blood volume and oxygenation.

Although some large vessels surrounded by CSF on the brain surface close to the activation area displayed NBS changes (see for instance three of them in Figure 3A, upper row), negative BOLD, in general, was more apparent in periventricular regions more distant from the activation area. An explanation of this finding might be related to the vessel size and distance from the activation area. The draining veins from active areas combine a negative volume effect with a positive BOLD effect, e.g., the increased flow in the activated area results in a lower deoxyhemoglobin concentration. Combined with the increase in volume, the total deoxyhemoglobin content change may be close to zero. The negative volume effect would be expected to be stronger farther from the activation site, where there could be volume changes without an appreciable decrease in deoxyhemoglobin concentration. Indeed, in the case of a very long vascular pathway, one would expect the venous oxygenation to be close to tissue oxygenation irrespective of flow velocity, because of oxygen diffusion across the vasculature. An activity induced flow velocity increase may therefore lead to a smaller decrease in deoxyhemoglobin concentration in largest veins compared with smaller veins. Moreover, as visible from our calculations (see Figure 6), larger baseline blood volume fractions induce larger signal decreases, thus negative BOLD is stronger for larger than for smaller vessels. For small oxygenation changes (e.g., dY/Y < 16.7), this effect is even more pronounced for vessels with their axis perpendicular (Figure 6B) rather than parallel (Figure 6A) to B₀.

Spurious negative activity is generally observed in scattered voxels throughout the brain, and their number may exceed that expected from the applied statistical threshold. In part, this activity may be caused by noise sources that do not obey model assumptions such as motion and cardiac and respiratory rate. However, these phenomena are unlikely to have caused the current observation for two reasons: first, anticorrelated activity was modulated by behavior, and this was not the case for motion-related parameters. Second, physiologic noise was removed before seed analysis, accounting both for signal fluctuations related to the phase of physiologic cycles (Glover et al, 2000) and to changes in physiologic rates (dual-lagged model; Bianciardi et al, 2009b). Periventricular activation was defined on the residual signal fluctuations after removal of these noise sources; therefore, the impact of physiologic noise on periventricular activity was negligible. Moreover, anticorrelations were not introduced by physiologic noise correction, as they were present both before and after physiologic noise correction (see Figure 2B). We also verified if the use of the dual-lagged model for fluctuations in physiologic rates optimized for the gray matter (Bianciardi et al, 2009b) could have biased the obtained result in periventricular areas. This was not the case, as RV and cardiac-rate regressors did not significantly correlate with signals in ROI_AC for any introduced temporal shift (lag times ranging between −21 and 21 seconds).

In this study at 7T (as in van der Zwaag et al, 2009), negative activity occurred not only in periventricular regions, as observed in previous work at lower field strength (Hong et al, 2009; Olbrich et al, 2009; Moosmann et al, 2003), but also in scattered sulcal areas with large pial venous vessels. This invites the question of why these sulcal effects were visible only at high magnetic fields. This could have a number of reasons, including the choice of experimental parameters. For example, the size of the effect is expected to depend on the spatial resolution, with a generally smaller effect expected for the relative coarse resolution used in typical fMRI experiments. Furthermore, at coarser resolution, negative effects caused by blood volume are more likely to be canceled out by positive tissue T*₂ effects. Finally, the size of the effect will depend on TR, TE, flip angle, and field strength used for fMRI acquisition, as they govern the relative size of the counteracting S₀ and T*₂ effects.

Conclusion

Data obtained in this study suggest that blood volume effects in large veins downstream from activated brain areas may reduce the BOLD signal and lead to apparent negative activation. The effect is expected to be most apparent in high-resolution experiments, where it may appear spatially distinct from positive activation and veins occupy a larger volume fraction in some voxels. In lower resolution experiments, negative activation may not be spatially distinct but rather reduce signal in positively activated regions. Conversely, the findings of this study suggest that increasing the resolution in fMRI may increase the sensitivity by reducing signal cancellation caused by concurrent T*₂ and blood volume effects.